Stability of implicit extrapolation methods. Abstract. Multilevel methods are generally based on a splitting of the solution

Size: px

Start display at page:

Download "Stability of implicit extrapolation methods. Abstract. Multilevel methods are generally based on a splitting of the solution"

Lora Fitzgerald
5 years ago
Views:

1 Contemporary Mathematics Volume 00, 0000 Stability of implicit extrapolation methods Abstract. Multilevel methods are generally based on a splitting of the solution space associated with a nested sequence of coarser grids. Besides the straightforward application of extrapolation on the grid system, we propose to use extrapolation implicitly, similar to multigrid -extrapolation. This implicit extrapolation, when applied to linear nite elements is related to the p-version of the nite element method. The method depends on a stability condition which must be enforced by a suitable modication of the problem. 1. Introduction Multilevel methods decompose the solution space in a nested sequence of subspaces. These subspaces are then used to construct a multigrid or multilevel preconditioned conjugate gradient method. Thus the multilevel structure is the basis for the ecient solution of a partial dierential equation (PDE). In this setting the solution is dened in topmost space and the multilevel structure is just used to accelerate some iterative solution method. Besides this algebraic perspective, the multilevel structure may be also used to improve the accuracy of the discretization itself. Under certain conditions, the nested mesh and space structure can be exploited by various extrapolation schemes. Extrapolation results for nite dierence methods can be found in [7] and for nite elements in [1]. All explicit extrapolation techniques rely on the existence of global error expansions for the approximate solutions, typically of the form (1.1) u h? u = kx i=1 h i e i + R k+1 ; where u h and u are the numerical and the true solution of the dierential equation, where e i are functions independent of h, and where R k+1 is a remainder term. The parameter h denotes the mesh width and can be interpreted as identifying a single space V h in the nested multilevel structure. Once an expansion 1991 Mathematics Subject Classication. 65N22, 65N50, 65N55. Key words and phrases. Implicit extrapolation, multigrid, multilevel subspace splitting. c0000 American Mathematical Society /00 $ $.25 per page 1

2 2 STABILITY OF IMPLICIT EXTRAPOLATION METHODS of the form (1.1) has been proven to exist, and when the exponent 1 of the leading term has been identied, then a linear combination of u h and u 2h of the form u h = ? 1 u 1 h? 2 1? 1 u 2h will lead to an approximation where the dominating error term is eliminated. Using this scheme recursively leads to the well-known Richardson extrapolation table. Extrapolation is only computationally feasible, when the coecients i are known, and when they form a quickly growing sequence. If the coecients are not known a priory, more complicated algorithms, like the -algorithm (see [3]) would have to be used. If the 1 do not grow quickly, so that the expansion has many terms of almost equal order, extrapolation will be less attractive, since many linear combinations are needed before the approximation order is signicantly increased. Unfortunately, these situations are typical in many practical PDE problems when reentrant corners, interfaces with jumps in the coecients or rough data may lead to singularities, each of which may create its own error contributions to the error expansion with a new set of exponents i. In the case of reentrant corner singularities when the i are still known, a local extrapolation technique has been proposed in [2], based on meshes rened locally according to the interior angles. In this paper we will discuss another variant of extrapolation, where the Richardson principle is applied indirectly. Such implicit extrapolation methods have been introduced in [9, 11, 5]. They are closely related to the so-called -extrapolation in multigrid methods, see e.g. [4, chap ]. Implicit extrapolation is based on a local element-by-element analysis of numerical quadrature and dierentiation rules. Thus they primarily depend on the smoothness of the shape functions rather than the global regularity of the solution. The regularity of the solution is of course required to justify the use of high order (polynomial) approximations in general. The advantage of implicit extrapolation is, that it is formally not applied to the solution itself, but only to the nite element functions approximating the solution. Thus the specic analysis is independent of the solution properties. Whether high order nite elements are suitable to approximate a given solution can be decided independently. Furthermore, implicit extrapolation is suitable for nonuniform grids as they may occur in a local renement context. Using extrapolation in a multilevel context is especially attractive, when the extrapolation to obtain higher order discretizations is eciently integrated with a fast multilevel solver. For integrating Richardson extrapolation with a full multigrid method and a comparison with multigrid -extrapolation. see e.g. Lin and Schuller [13]. The implicit extrapolation which is the main topic of this paper is most eciently implemented by the multigrid -extrapolation algorithm. Due to the space limitations in this paper we refer here for all details of the al-

3 STABILITY OF IMPLICIT EXTRAPOLATION METHODS 3 gorithms to [9, 5]. Here we remark only that implicit extrapolation amounts to only a trivial change of a conventional (FAS) multigrid method, which consists of only a multiplication by an additional extrapolation factor in the ne-to-coarse transfer. Thus integrating implicit extrapolation into an existing multilevel algorithm is easy, and the basic solver remains untouched, since the higher order accuracy is obtained by a defect correction-like iteration. 2. The implicit extrapolation method For exposition, we consider the simplest case of an elliptic PDE in one dimension only (2.1) u 00 = f; u(0) = u(1) = 0: We will use the equivalent formulation as a minimization problem (2.2) min E(u); where E(u) = u2h0 1(0;1) Z 1 0 (u 0 (x)) 2? 2u(x)f(x) dx: Next, we introduce the mesh 0 = x 0 < x 1 < : : : < x n = 1, and discretize (2.2) directly, by representing the continuous function u by the vector u h = (u h 0; : : : ; u h n) T. With the mesh widths h i = x i+1?x i and the midpoints x i+1=2 = 1=2(x i+1 + x i ), we may replace u 0 (x i+1=2 ) u(x i+1)? u(x i ) h i and the integration by the midpoint sum Z 1 0 f(x) dx n?1 X i=0 h i f(x i+1=2 ) for the numerical approximation to (u 0 (x)) 2 and the trapezoidal sum Z 1 0 f(x) dx n?1 X i=0 h i 2 (f(x i+1) + f(x i )) for approximating the integral over u(x)f(x), see Fig. 1. Figure 1. Trapezoidal and Midpoint rule for integration

4 4 STABILITY OF IMPLICIT EXTRAPOLATION METHODS These approximations combined, lead to E(u) E h (u h ) = " n?1 X 1 ui+1? u i h i 2 h i i=0 The corresponding normal equations are?u i+1 + u i h i+1 + u i? u i?1 h i u 0 = f i+1u i+1 + f i u i + h i+1 + h i 2 f i = 0 for i = 1; : : : ; N? 1 u N = 0 : Thus we have recovered the discretization by second order nite dierences, or, equivalently, the discretization by piecewise linear nite elements with lumped mass matrix. The basic reason for deriving the discrete system in this form is the existence of asymptotic error expansions for (2.3) E(u)? E h (u) = e 2 h 2 + : : : + e 2k h 2k + R 2k+1 when u is suciently smooth. This result holds even for a nonuniform basic discretization x 0 ; : : : ; x n, if only the rened grids are constructed by recursively inserting the interval midpoints. This result is proved in [6]. Based on the expansion (2.3) we may now consider extrapolated functionals of the form (2.4) E h (u h ) = 4 3 E h(u h ) 1? 3 E 2h(u h ); Eh (u h ) = 16 E 15 h (u h ) 1? E 15 2h (u h ); etc. Here we interpret u h simply as a vector of values, and note that E 2h (u h ) only uses every second value in u h. Obviously, each of the extrapolated functionals denes a new system of normal equations for u h. In section 3 we will give conditions, when the higher order representation of the functional results in an improved accuracy for u h. Before we continue this argument, we mention that results analogous to (2.3) hold in two space dimensions, as proved in [12] for the case of triangular meshes. Of course this is a crucial result, since the main interest here is in methods which generalize to higher space dimensions. In [5] it is furthermore shown that the implicit extrapolation method is equivalent to using higher order nite elements in a special case. 3. Stability of implicit extrapolation In the above section we have introduced the implicit extrapolation principle for deriving discretizations of potentially higher order for elliptic PDE. Here we will now interpret the technique in a nite element setting. Thus we interpret u h as a nite element approximation with lowest oder, that is piecewise linear nite elements. By the extrapolation, as in (2.4), we construct higher order representations of the integrals. These extrapolated quadrature rules are dened for 2 # :

5 STABILITY OF IMPLICIT EXTRAPOLATION METHODS 5 the continued renement of a basic mesh x 0 ; x 1 ; : : : ; x n by recursively inserting the interval midpoints. If the basic mesh is associated with the space of piecewise linear nite element functions V h, it would seem natural to identify each of the rened meshes with the corresponding (hierarchical) nite element spaces V h V h=2 V h=4 : : :. of piecewise linear functions. However, since the basic integration/dierentiation rules are already correct in all these spaces, their approximation properties cannot be improved by their extrapolation. If there is a positive eect of the implicit extrapolation, it cannot be directly understood within the h-renement space structure. Therefore, we consider a p-renement where the basic space of piecewise linear approximations V h is enlarged by piecewise polynomials of increasing order. Thus the rst renement from space V h to Vh=2 with a rst step of extrapolation corresponds to adding quadratic functions in each interval (x i ; x i+1 ). The next step to Vh=4 introduces two additional degrees of freedom in each element and this corresponds to adding cubic and 4th order basis functions. In the next level V h=8 another four degrees of freedom are added and thus all polynomials up to degree 8. At the same time as we add p-renement, the extrapolation increases the accuracy of the integration. Unfortunately, this increase of accuracy does not keep up with the rapid growth of the paces. Only in the rst step, when quadratic shape functions are added, the extrapolation provides suciently accurate integration rules. In the second step of extrapolation only the functions up to degree 3 are integrated exactly, however, cubic and 4th order functions are added. Thus there is a mismatch between the functions in the nite element space and the accuracy of the extrapolated quadrature rules. Only in the rst step the method (as dened so far) is equivalent to a p-version nite element method. In conventional nite element analysis, the need for numerical quadrature to compute the stiness matrix (and right hand side) is rarely made explicit. The nite element space is basic, and the use of another space (say of nodal values) for numerical quadrature is usually not explicit in the basic analysis. Here, in the context of implicit extrapolation, the converse is true. We directly work in the space used for quadrature and our interpretation of this as a p-version nite element space is just an artifact. Generally speaking, we are faced with the following situation. The continued renement has produced a large space V which we deliberately identify with a p-renement of our basic discretization. The associated nite element problem is denoted by (3.1) mine(u); where E(u) = a(u; u)? 2(f; u): u2v In the space V, however, we cannot assume that we represent the functionals correctly. Our extrapolation procedure falls behind in accuracy,

6 6 STABILITY OF IMPLICIT EXTRAPOLATION METHODS so what we really solve is not (3.1) but (3.2) min ~u2v ~E(~u); where ~ E(~u) = ~a(~u; ~u)? 2(f; ~u) Though we cannot say, how accurate ~ E is with respect to E in the full space V, we have constructed the method such, that V has a subspace W, where both functionals agree well with each other, say (3.3) je(w)? ~ E(w)j jjwjj2 for all w 2 W : In W we consider the equation dened by the original functional (3.4) min E(w) w2w The best we can hope for in this situation is that our computed solution ~u of problem (3.2) is close to the solution w of problem (3.4). We have the following theorem. Theorem 3.1 (Stability of implicit extrapolation). Let V be a Hilbert space with norm jj jj and bilinear forms a(; ) and ~a(; ), f 2 V with positive constants c 1 ; c 2 such that c 1 jjvjj 2 a(u; u) c 2 jjvjj 2 : Let u, ~u, and w be dened by ( 3.1), ( 3.2), and ( 3.4), respectively. Assume that ( 3.3) holds and that (3.5) ~a(v; v) a(v; v) for all v 2 V: Then there exists c > 0 such that jjw? ~ujj c( p jjwjj + jjw? ujj): Proof. ~a(w? ~u; w? ~u) = ~a(w; w? ~u)? ~a(~u; w? ~u) = ~a(w; w? ~u)? a(u; w? ~u) = ~a(w; w? ~u)? a(w; w? ~u) + a(w? u; w? ~u) = (~a? a)(w; w? ~u) + a(w? u; w? ~u) p p p (~a? a)(w; w) (~a? a)(w? ~u; w? p ~u) + a(w? u; w? u) a(w? ~u; w? ~u) (~a? a)(w; w) ~a(w? ~u; w? ~u) + a(w? u; w? u) ~a(w? ~u; w? ~u): Therefore p~a(w? ~u; w? ~u) p (~a? a)(w; w) + p a(w? u; w? u): and thus there exists c > 0 with jjw? ~ujj c( p jjwjj + jjw? ujj):

7 STABILITY OF IMPLICIT EXTRAPOLATION METHODS 7 This stability theorem requires condition (3.5), which states that we must not underestimate the true energy E(u) in our numerical analogue ~ E(u). In many practical situations this condition is violated as soon as we attempt a second step of extrapolation, and the potential increase in accuracy is not reected in the solution. A computational remedy has been suggested in [10]. Obviously it is sucient so solve (3.2) subject to the constraint ~u 2 W. If such a constraint is used to make V coincide with W, the instability cannot occur. However, condition (3.5) indicates that we need not strictly impose ~u 2 W, but it suces to correct a possible underestimate of the energy of solution components which do not lie in W. Since the orthogonal complement of W in V consists of all polynomials of degree larger than some k 0 and smaller than k 1, it is simple to construct such a correction. Thus (3.2) becomes modied to (3.6) min ~u2v ~E(~u) + D(~u); where is a suciently large real parameter and D(~u) is a quadratic form which vanishes on all (piecewise) polynomials of degree k 0 and is positive for all (piecewise) polynomials of degree > k 0 and k 1. Thus the accuracy condition (3.3) is maintained, and clearly, for some large enough, the stability condition (3.5) will be satised, too. Consequently, an ansatz of the form (3.6) where E ~ is constructed by implicit extrapolation may obtain any approximation order, provided that the penalty term D is chosen appropriately. D(~u) can e.g. be constructed by computing the nite dierences estimating the kth derivatives for k 0 < k k 1 in each element. Thus D(~u) is a local operator, eecting only nodal values within a single element. Computational experiments showing the eectiveness of this technique are given in [10]. 4. A numerical experiment for a singular interface problem The implicit extrapolation technique can be generalized to two and three space dimensions. We will now present a two-dimensional numerical experiment in a nontrivial situation. We consider the diusion equation (4.1) (4.2) r (x; y)ru(x; y) = 0 in (0; 1) 2 u(x; y) = g(x; y) 1) 2 ; where the coecient has jumps across the lines x = 1=2 and y = 1=2. particular, we assume In (x; y) = 1 in (0; 1=2) (1=2; 1) [ (1=2; 1) (0; 1=2) 3 in (0; 1=2) 2 [ (1=2; 1) 2 : We choose the boundary data such that the solution has the form u(x; y) = u(r; ) = ()r (2=3) ;

8 8 STABILITY OF IMPLICIT EXTRAPOLATION METHODS where (r; ) are polar coordinates with respect to the point (x; y) = (1=2; 1=2), and where () is of the form () = sin(2=3 + ) in each of the four quadrants of the domain, with chosen such that the interface conditions for the normal derivatives are satised. The solution is depicted in Fig. 2. Since the singularity is of the same type as in the case of an L-shaped Figure 2. Solution of problem (4.1,4.2) and convergence rates. domain, we expect poor convergence. It has been shown, that in a point-wise sense and for a uniform discretization with piecewise linear triangular elements, we can only expect a convergence rate of O(h 4=3 ) This is reected in the upper (unbroken) line of Fig. 2. A conventional technique to improve on this is to use local renement to better resolve the singularity. An alternative is based on extrapolation, as in [2] or [8]. This is the approach we will take, focussing on implicit extrapolation. Thus we solve a system of the form (2.3), however, we must adapt the extrapolation parameter to the nonstandard leading term in the expansion. We thus minimize (2 4=3?1)?1 (2 (4=3) E h (u h )?E 2h (u h )). The convergence rate for the corresponding solution is shown in the lower (broken) line of Fig. 2. The errors visualized in Figs. 3. The eect of implicit extrapolation is clearly visible. Since we have not rened locally, the poor resolution of the singularity is not signicantly improved. However, with implicit extrapolation, the global spreading of the error, the so-called pollution eect is fully suppressed. This is also clear form the convergence graph in Fig. 2, where it is shown that implicit extrapolation has recovered O(h 2 ) convergence away from the singularity. 5. Conclusions In this paper we have presented briey the implicit extrapolation method. We have proved a stability condition and have suggested computational techniques

9 STABILITY OF IMPLICIT EXTRAPOLATION METHODS 9 Figure 3. Error for standard discretization and implicit extrapolation. to satisfy the stability condition in a general setting. A numerical example has shown, how the method can be used even in the case of singular interface problems. Many more aspects of the method, in particular its combination with splitting extrapolation (see [14]) and the so-called sparse grid techniques remain to be analyzed in the future. Some preliminary results are given in [10]. References 1. H. Blum, Q. Lin, and R. Rannacher, Asymptotic error expansions and Richardson extrapolation for linear nite elements, Numer. Math. 49 (1986), 11{ H. Blum and R. Rannacher, Extrapolation techniques for reducing the pollution eect of reentrant corners in the nite element method, Numer. Math. 52 (1988), 539{ C. Brezinski and M. Redivo Zaglia, Extrapolation methods: Theory and practice, Studies in Computational Mathematics, vol. 2, North Holland, Amsterdam, W. Hackbusch, Multigrid methods and applications, Springer Verlag, Berlin, M. Jung and U. Rude, Implicit extrapolation methods for multilevel nite element computations, Preliminary Proceedings of the Colorado Conference on Iterative Methods, Breckenridge, Colorado, April 4-10, 1994 (T. Manteuel, ed.), 1994, To appear in S.J.Sci.Comput., 17(1), J.N. Lyness, The calculation of the Stieltjes' integral, Num. Math. 12 (1968), 252{ G.I. Marchuk and V.V. Shaidurov, Dierence methods and their extrapolations, Springer, New York, U. Rude, On the accurate computation of singular solutions of Laplace's and Poisson's equation, Multigrid Methods: Theory, Applications, Supercomputing: Proceedings of the Third Copper Mountain Conference on Multigrid Methods, April 5-10, 1987 (New York, NY, USA) (S. F. McCormick, ed.), Marcel Dekker, , Adaptive higher order multigrid methods, Proceedings of the Third European Conference on Multigrid Methods, October 1-4, 1990 (Boston, MA, USA) (W. Hackbusch and U. Trottenberg, eds.), Birkhauser, 1991, International Series of Numerical Mathematics, Vol. 98, pp. 339{ , Extrapolation and related techniques for solving elliptic equations, Bericht I-9135, Institut fur Informatik, TU Munchen, September , The hierarchical basis extrapolation method, SIAM J. Sci. Stat. Comput. 13 (1992), no. 1, 307{318, Proceedings of the First Copper Mountain Conference on Iterative Methods, April 1-5, 1990, T. Manteuel ed.

10 10 STABILITY OF IMPLICIT EXTRAPOLATION METHODS 12., Extrapolation techniques for constructing higher order nite element methods, Bericht I-9304, Institut fur Informatik, TU Munchen, A. Schuller and Qun Lin, Ecient high order algorithms for elliptic boundary value problems combining full multigrid techniques and extrapolation methods, Arbeitspapiere der GMD 192, Gesellschaft fur Mathematik und Datenverarbeitung, December Lu Tao, Tsi min Shih, and Chin bo Liem, An analysis of splitting extrapolation for multidimensional problems, Systems Science and Mathematical Sciences 3 (1990), no. 3, 261{272. (U. Rude) Institut fur Informatik Technische Universitat Munchen Arcisstr. 21, D Munchen 2, Germany address: ruede@informatik.tu-muenchen.de

A MULTIGRID ALGORITHM FOR. Richard E. Ewing and Jian Shen. Institute for Scientic Computation. Texas A&M University. College Station, Texas SUMMARY

A MULTIGRID ALGORITHM FOR. Richard E. Ewing and Jian Shen. Institute for Scientic Computation. Texas A&M University. College Station, Texas SUMMARY A MULTIGRID ALGORITHM FOR THE CELL-CENTERED FINITE DIFFERENCE SCHEME Richard E. Ewing and Jian Shen Institute for Scientic Computation Texas A&M University College Station, Texas SUMMARY In this article,